Divide-and-conquer method and system for the detection of lung nodule in radiological images

1. A method of identifying abnormalities in radiological images of the chest, the method comprising the steps of: receiving a digital chest image; discarding pixels outside of the lung zone; dividing the lung zone into different overlapped zones based on the geometry and subtlety distribution and image characteristics of abnormalities and normal anatomic structure in each zone; enhancing abnormality signals by using enhancement techniques specifically developed for each zone; preliminarily selecting suspect abnormalities in each zone; extracting image features in each zone; classifying abnormalities and normal anatomic structures in each zone; integrating outputs from different zones; and providing an indication of abnormalities.

2. A method according to claim 1 , wherein the step of dividing further comprises the steps of: extracting a non-obscured lung zone using a neural network; and obtaining an obscured lung zone by excluding from the lung zone the pixels of the image corresponding to the non-obscured lung zone.

3. A method according to claim 2 , wherein said different overlapped zones comprise: a zone of obscured lung; a zone of clavicle; a zone of peripheral lung edge; a zone of peripheral lung central; a zone of hilum; and a zone of diaphragm.

4. A method according to claim 3 , wherein said zone of obscured lung includes spine, heart, and mediastinum.

5. A method according to claim 3 , wherein said zone of clavicle comprises an area extending from the tops of right and left portions of said non-obscured lung zone and covering a region comprising a predetermined portion of the vertical distance toward the bottoms of said right and left portions of said non-obscured lung zones.

6. A method according to claim 3 , wherein said zone of peripheral lung edge is obtained by the step of: performing image processing methods including morphological processing with a structure element along the contour of the ribcage.

7. A method according to claim 6 , wherein said morphological processing comprises erosion.

8. A method according to claim 3 , wherein said zone of peripheral lung central is obtained by the step of: performing image processing methods including morphological processing with two structure elements along the contour of the ribcage.

9. A method according to claim 8 , wherein said morphological processing comprises erosion.

10. A method according to claim 3 , wherein said zone of hilum is obtained by the step of: performing image processing methods including morphological processing with two structure elements along the contour of the ribcage; and performing image processing methods including morphological processing with a structure element along the contour of the heart and spine.

11. A method according to claim 10 , wherein at least one of said morphological processings comprises erosion.

12. A method according to claim 3 , wherein said zone of the diaphragm is obtained by the step of: performing image processing methods including morphological processing with a structure element along the diaphragm.

13. A method according to claim 12 , wherein said morphological processing comprises erosion.

14. A method according to claim 3 , wherein said step of enhancing abnormality signals further comprises the step of: applying specific image enhancement methods tailored to each zone independently to suppress background structures and to further enhance abnormality-to-background contrast.

15. A method according to claim 14 , wherein said image enhancement method tailored to said zone of obscured lung comprises the step of: using contrast enhancement methods, to obtain a more uniform histogram of the zone and enhance the contrast of obscured lung.

16. A method according to claim 14 , wherein said image enhancement method tailored to said zone of clavicle comprises the steps of: performing edge detection methods; and using an edge removal method.

17. A method according to claim 16 , wherein said edge detection methods include a Hough transform.

18. A method according to claim 14 , wherein said image enhancement method tailored to said zone of peripheral lung comprises the step of: using edge detection and removal methods to suppress the ribcage structure.

19. A method according to claim 18 , wherein said edge detection and removal methods include a Hough transform.

20. A method according to claim 14 , wherein said image enhancement method tailored to said zone of peripheral lung central comprises the step of: performing image subtraction using median and matched filtering by spatially convolving the pixels of the zone with multiple sphere profiles of different size nodule images.

21. A method according to claim 14 , wherein said image enhancement method tailored to said zone of hilum comprises the step of: performing image subtraction using median and matched filtering by spatially convolving the pixels of the zone with multiple synthetic nodule profiles of different sizes.

22. A method according to claim 14 , wherein said image enhancement method tailored to said zone of diaphragm comprises the step of: increasing the abnormality-to-background contrast using an inverse contrast ratio mapping technique.

23. A method according to claim 3 , wherein said step of preliminarily selecting suspect abnormality areas further comprises the steps of: clustering the suspect abnormality areas into different groups; and using different threshold and sensitivity and specificity values in each different zone.

24. A method according to claim 23 , wherein said step of preliminarily selecting suspect abnormality areas in said zone of obscured lung is performed with high specificity.

25. A method according to claim 23 , wherein said step of preliminarily selecting suspect abnormality areas in said zone of clavicle is performed with high sensitivity.

26. A method according to claim 23 , wherein said step of preliminarily selecting suspect abnormality areas in said zone of peripheral lung edge is performed with high sensitivity.

27. A method according to claim 23 , wherein said step of preliminarily selecting suspect abnormality areas in said zone of peripheral lung central is performed with high sensitivity.

28. A method according to claim 23 , wherein said step of preliminarily selecting suspect abnormality areas in said zone of hilum is performed with high specificity.

29. A method according to claim 23 , wherein said step of preliminarily selecting suspect abnormality areas in said zone of diaphragm is performed with high specificity.

30. A method according to claim 3 , wherein said step of classifying includes the steps of: training specific classifiers using the selected suspects and abnormality in each zone to classify an abnormality and normal anatomic structures in each zone; and training at least one neural network by using cross-validation using SUB-A z techniques.

31. A method according to claim 30 , wherein said step of classifying further comprises the step of, for said zone of obscured lung, using a back-propagation neural network.

32. A method according to claim 30 , wherein said step of classifying further comprises the step of, for said zone of clavicle, comparing the density around the clavicle bone area of both the left and right lungs to detect abnormalities.

33. A method according to claim 30 , wherein said step of classifying further comprises the step of, for said zone of peripheral lung edge, training a back-propagation neural network.

34. A method according to claim 30 , wherein said step of classifying further comprises the step of, for said zone of peripheral lung central, training a back-propagation neural network using those of said selected suspects located in said zone of peripheral central.

35. A method according to claim 30 , wherein said step of classifying further comprises the step of, for said zone of hilum, using a convolution neural network.

36. A method according to claim 30 , wherein said step of classifying further comprises the step of, for said zone of diaphragm, thresholding a matching filter.

37. A method according to claim 30 , wherein a desired performance of each of said zone-based classifiers is obtained by training each zone classifier to have sensitivity and specificity performance according to the particular zone.

38. A method according to claim 30 , wherein said cross-validation using SUB-A z is used to cross-validate the classifier for each zone.

39. A method according to claim 1 , wherein said step of integrating outputs comprises the step of combining the outputs from different zones based on the sensitivity and specificity performance for each zone in said step of classifying.

40. A method according to claim 1 , further comprising the step of: utilizing a parallel processing engine for processing each zone, said parallel processing engine processing each zone independently and in parallel.

41. A method according to claim 3 , wherein said step of extracting image features in each zone comprises the step of, for the zone of obscured lung, extracting size, circularity and elongation for each suspected abnormality area in that zone.

42. A method according to claim 41 , where said step of extracting size, circularity and elongation comprises the step of employing a bit quads method to extract size.

43. A method according to claim 3 , wherein said step of extracting image features in each zone comprises the step of, for the zone of clavicle, extracting gradient features for each suspected abnormality area in that zone.

44. A method according to claim 43 , where said gradient features include amplitude and orientation edge maps.

45. A method according to claim 3 , wherein said step of extracting image features in each zone comprises the step of, for the zone of peripheral lung edge, extracting circularity, area and moment invariants for each suspected abnormality area in that zone.

46. A method according to claim 3 , wherein said step of extracting image features in each zone comprises the step of, for the zone of peripheral lung central, extracting circularity, size and amplitude and orientation edge maps for each suspected abnormality area in that zone.

47. A method according to claim 3 , wherein said step of extracting image features in each zone comprises the step of, for the zone of hilum, normalizing image pixels of each suspected abnormality area in that zone.

48. A method according to claim 47 , wherein said step of normalizing image pixels comprises the steps of: computing an average pixel value, 0 , of the 8-connected pixels surrounding each pixel of a suspected abnormality area; and for each pixel value x, mapping the pixel value to a new pixel value, x*, by using the equation, x * = 1 1 + - ( x - 0 ) where is a decaying control parameter.

49. A method according to claim 3 , wherein the step of extracting image features in each zone comprises, for the zone of diaphragm, extracting circularity and size for each suspected abnormality area in that zone.

50. A system for identifying abnormalities in radiological images of the chest, comprising: (a) an image input unit, said image input unit receiving a digital chest image; (b) a memory unit for storing at least one digital chest image; (c) an image processing unit that detects said abnormalities, the image processing unit comprising: an image zoning unit that divides said digital chest image into overlapped zones, thereby generating a zoned image; a zone-based image enhancement unit that enhances each zone of said zoned image using an enhancement technique tailored to that zone, thereby generating a zone-enhanced image; a zone-based suspect selection unit that processes each zone of said zone-enhanced image to extract zone-grouped suspect abnormality areas; a zone-based feature extraction unit that processes said zone-grouped abnormality areas of each zone using a technique tailored to that particular zone, thereby generating zone-grouped suspect features; a zone-based classification unit that uses different classifiers tailored to the different zones to process said zone-grouped suspect features and to thereby generate suspect classification scores corresponding to said zone-grouped suspect abnormality areas; and a zone-based data fusion unit that processes said suspect classification scores associated with said zone-grouped suspect abnormality areas to provide final abnormality areas; and (d) an image output unit.

51. A system according to claim 50 , wherein at least one of said zone-based image enhancement unit, said zone-based suspect selection unit, said zone-based feature extraction unit, and said zone-based classification unit comprises components that implement parallel processing of the various zones.

52. A system according to claim 50 , wherein said image zoning unit discards pixels of said digital chest image located outside the lung region; further divides the lung region into obscured and non-obscured lung zones; and sub-divides the non-obscured lung zone into further zones based on the geometry and subtlety distribution and image characteristics of abnormalities in each of the zones.

53. A system according to claim 52 , wherein said further zones include the following: obscured lung, clavicle, peripheral lung edge, peripheral lung central, hilum, and diaphragm.

54. A system according to claim 53 , wherein said obscured lung zone includes spine, heart, and mediastinum.

55. A system according to claim 53 , wherein said clavicle zone comprises an area extending from the tops of right and left portions of said non-obscured lung zone and covering a region comprising a certain portion of the vertical distance toward the bottoms of said right and left portions of said non-obscured lung zones.

56. A system according to claim 53 , wherein said image zoning unit obtains said peripheral lung edge zone by performing image processing methods including morphological processing with a structure element along the contour of the ribcage.

57. A system according to claim 56 , wherein said morphological processing comprises erosion.

58. A system according to claim 53 , wherein said image zoning unit obtains said peripheral lung central zone by performing image processing methods including morphological processing with two structure elements along the contour of the ribcage.

59. A system according to claim 58 , wherein said morphological processing comprises erosion.

60. A system according to claim 53 , wherein said image zoning unit obtains said hilum zone by performing image processing methods including morphological processing with two structure elements along the contour of the ribcage.

61. A system according to claim 60 , wherein said morphological processing comprises erosion.

62. A system according to claim 53 , wherein said image zoning unit obtains said diaphragm zone by performing image processing methods including morphological processing with a structure element along the diaphragm.

63. A system according to claim 62 , wherein said morphological processing comprises erosion.

64. A system according to claim 53 , wherein said zone-based image enhancement unit implements specific image enhancement methods tailored to each zone independently to suppress background structures and to further enhance abnormality-to-background contrast.

65. A system according to claim 64 , wherein said zone-based image enhancement unit comprises zone-specific image enhancement units, each of which implements one of said image enhancement methods.

66. A system according to claim 64 , wherein the image enhancement method tailored to said obscured lung zone uses contrast enhancement methods to obtain a more uniform histogram and enhance the contrast of the obscured lung zone.

67. A system according to claim 66 , wherein said contrast enhancement methods include histogram equalization.

68. A system according to claim 64 , wherein the image enhancement method tailored to said clavicle zone includes performing edge detection methods and using an edge removal technique.

69. A system according to claim 68 , wherein said edge detection methods include a Hough transform.

70. A system according to claim 64 , wherein the image enhancement method tailored to said peripheral lung zone uses edge detection and removal methods to suppress the ribcage structure.

71. A system according to claim 70 , wherein said edge detection and removal methods include a Hough transform.

72. A system according to claim 64 , wherein the image enhancement method tailored to said peripheral lung central zone performs median and matched filtering by spatially convolving the pixels of the zone with multiple synthetic sphere profiles of different sizes of nodule images.

73. A system according to claim 64 , wherein the image enhancement method tailored to said hilum zone performs matched filtering by spatially convolving the pixels of the zone with multiple sizes of synthetic nodule profiles.

74. A system according to claim 64 , wherein the image enhancement method tailored to said diaphragm zone increases the abnormality-to-background contrast using an inverse contrast ratio mapping technique.

75. A system according to claim 53 , wherein said zone-based suspect selection unit clusters suspect abnormality areas into different groups and uses different threshold and sensitivity and specificity values in each different zone.

76. A system according to claim 75 , wherein in said obscured lung zone high specificity is used.

77. A system according to claim 75 , wherein in said clavicle zone high sensitivity is used.

78. A system according to claim 75 , wherein in said peripheral lung edge zone high sensitivity is used.

79. A system according to claim 75 , wherein in said peripheral lung central zone high sensitivity is used.

80. A system according to claim 75 , wherein in said hilum zone high specificity is used.

81. A system according to claim 75 , wherein in said diaphragm zone high specificity is used.

82. A system according to claim 53 , said zone-based classification unit comprising specific classifiers corresponding to each of said zones, each of said classifiers being trained using selected suspect and abnormality data of a particular zone, to enable that classifier to classify an abnormality and normal anatomic structures in the particular zone; and wherein at least one of said specific classifiers comprises a neural network, said neural network being trained by using cross-validation using SUB-A z techniques.

83. A system according to claim 82 , wherein the classifier corresponding to the obscured lung zone comprises a back-propagation neural network.

84. A system according to claim 82 , wherein the classifier corresponding to the clavicle zone comprises means for comparing the density around the clavicle bone area of both the left and right lungs to detect abnormalities.

85. A system according to claim 82 , wherein the classifier corresponding to the peripheral lung edge zone comprises a back-propagation neural network.

86. A system according to claim 82 , wherein the classifier corresponding to the peripheral lung central zone comprises a back-propagation neural network.

87. A system according to claim 82 , wherein the classifier corresponding to the hilum zone comprises a convolution neural network.

88. A system according to claim 82 , wherein the classifier corresponding to the diaphragm zone thresholds a matching filter.

89. A system according to claim 82 , wherein a desired performance of each of said classifiers is obtained by training each classifier to have a sensitivity and specificity performance according to the particular zone.

90. A system according to claim 82 , wherein cross-validation using SUB-A z is used to cross-validate the classifier for each zone.

91. A system according to claim 82 , wherein said zone-based data fusion unit comprises means for combining the outputs from different zones based on sensitivity and specificity performance for each zone's classifier.

92. A system according to claim 50 , said image processing unit comprising a parallel processing engine for processing each zone, said parallel processing engine processing each zone independently and in parallel.

93. A system according to claim 92 , said parallel processing engine encompassing said zone-based image enhancement unit, said zone-based suspect selection unit, said zone-based feature extraction unit, and said zone-based classification unit.

94. A system according to claim 53 , wherein said zone-based feature extraction unit comprises means, for the obscured lung zone, for extracting size, circularity and elongation for each suspected abnormality area in that zone.

95. A system according to claim 94 , where said means for extracting size, circularity and elongation employs a bit quads method to extract size.

96. A system according to claim 53 , wherein said zone-based feature extraction unit comprises means, for the clavicle zone, for extracting gradient features for each suspected abnormality area in that zone.

97. A system according to claim 96 , where said gradient features include amplitude and orientation edge maps.

98. A system according to claim 53 , wherein said zone-based feature extraction unit comprises means, for the peripheral lung edge zone, for extracting circularity, area and moment invariants for each suspected abnormality area in that zone.

99. A system according to claim 53 , wherein said zone-based feature extraction unit comprises means, for the peripheral lung central zone, for extracting circularity, size and amplitude and orientation edge maps for each suspected abnormality area in that zone.

100. A system according to claim 53 , wherein said zone-based feature extraction unit comprises means, for the hilum zone, for normalizing image pixels of each suspected abnormality area in that zone.

101. A system according to claim 100 , wherein said means for normalizing image pixels comprises: means for computing an average pixel value, 0 , of the 8-connected pixels surrounding each pixel of a suspected abnormality area and, for each pixel value x, for mapping the pixel value to a new pixel value, x*, by using the equation, x * = 1 1 + - ( x - 0 ) where is a decaying control parameter.

102. A system according to claim 53 , wherein the zone-based feature extraction unit comprises, for the diaphragm zone, means for extracting circularity and size for each suspected abnormality area in that zone.